Amazon: AWS Is Accelerating. The AI Capex Bill Is Accelerating Faster.

The Kapital · US Stock Analysis · 20 August 2026

Amazon’s second quarter contains almost everything investors love and fear about the current AI cycle. Net sales rose 20% to $200.6 billion. AWS revenue jumped 37% to $42.2 billion and produced $16.6 billion of operating income. At the same time, trailing-twelve-month free cash flow fell to negative $7.6 billion as capital expenditures surged, primarily for AI infrastructure. Amazon is showing that demand is real. It is also showing how expensive it has become to satisfy that demand.

I increasingly think Amazon should be analyzed as three companies sharing one capital allocation system. The retail and logistics network is a scale machine with thin margins but enormous customer reach. Advertising monetizes that reach at attractive incremental economics. AWS turns computing infrastructure into a high-margin service business and is now carrying much of the AI investment burden.

The second quarter makes the tension between those pieces unusually clear. The company is generating more operating cash than ever, but it is spending even faster because management believes the next generation of cloud infrastructure requires a historic buildout.

That can be exactly the right decision. It can also be the point at which a great business begins earning lower returns on the next dollar of capital than on the last one.

AWS growth changed the tone of the quarter

AWS generated $42.2 billion of second-quarter revenue, up 37% year over year. Operating income reached $16.6 billion. Those numbers matter because cloud growth had been one of the central debates around hyperscaler capex. If companies are spending tens of billions on data centers while cloud growth slows, the market should become skeptical. If cloud growth accelerates while capacity remains constrained, aggressive capex becomes easier to defend.

For now, Amazon is in the second category.

At an annualized run rate above $160 billion, AWS is already one of the largest technology businesses in the world. A 37% growth rate at that scale adds more than $40 billion of annualized revenue if sustained. That is why I resist simplistic claims that the AI infrastructure cycle must immediately collapse because spending is high. The revenue base absorbing the spending is also enormous.

The operating margin inside AWS remains the crown jewel

What makes AWS so valuable is not just growth. It is the amount of profit attached to that growth. More than $16 billion of quarterly operating income from $42 billion of revenue implies an operating margin close to 40%.

That profitability gives Amazon a unique ability to fund infrastructure from within the business. A weaker cloud company would need external capital to build at this speed. Amazon can use AWS profits, retail cash flow and access to very deep capital markets to expand capacity ahead of demand.

The danger is that investors may extrapolate today’s AWS margin into a future where AI compute becomes more capital intensive and competition becomes more aggressive. GPUs, custom accelerators, networking gear, power and cooling all require large upfront spending. The depreciation expense arrives later. A period of very high capex can therefore make current operating margins look better than the long-run economics of newly built infrastructure.

I want to see whether AWS margins remain durable after the new AI asset base is fully reflected in depreciation.

Free cash flow is the uncomfortable part

Amazon said trailing-twelve-month operating cash flow rose 33% to $161.4 billion. That is an extraordinary amount of cash generation. Yet free cash flow fell to negative $7.6 billion because purchases of property and equipment increased by roughly $66 billion, primarily reflecting investments in artificial intelligence.

This is the most important financial sentence in the quarter.

A company can report rapidly growing operating income and still consume cash if investment requirements rise faster. That does not mean the investment is bad. It means valuation must include the return on that investment.

If Amazon spends an additional $60 billion and ultimately creates $15 billion of recurring incremental after-tax operating profit, the economics are excellent. If the same spending produces only $3 billion, shareholders would have been better served by a lower investment rate.

We cannot know the final answer today. But we can identify the metric that matters: incremental profit per incremental dollar of AI capital.

The Anthropic gain makes GAAP net income less useful

Amazon reported second-quarter net income of $62.6 billion. That number is real under accounting rules, but it is not a useful representation of recurring operating earnings. Reuters noted that the quarter included approximately $53.4 billion of pre-tax other income, primarily related to Amazon’s investment in Anthropic.

This is a perfect example of why headline earnings can mislead during a private-AI valuation boom. Mark-to-market gains can create enormous reported profits without creating equivalent operating cash flow. They can also reverse.

I treat the Anthropic stake as an asset with strategic and financial value, but I do not capitalize unrealized gains as if they were recurring earnings. For my valuation, AWS, advertising, retail margins and long-term cash generation matter much more.

Anthropic still matters strategically

Ignoring the accounting gain does not mean the relationship is irrelevant. Amazon’s investment in Anthropic can strengthen AWS by making the cloud an important home for frontier-model training and inference. It can also increase demand for Amazon’s custom Trainium and Inferentia chips.

This is where the ecosystem becomes interesting. Amazon is simultaneously a cloud provider, infrastructure financier, chip designer and investor in a major model company. If the pieces reinforce each other, Amazon can capture economics at several layers of the AI stack.

The concern is circularity. Amazon invests in an AI company, the AI company spends on AWS, AWS uses the revenue to justify more infrastructure, and Amazon invests again. The economic loop is healthy only if end customers ultimately pay enough for AI products to support the chain.

That is why I focus on external AWS demand, not merely affiliated workloads.

Retail is becoming a better business than the old narrative suggests

The common Amazon story still describes retail as a low-margin engine that exists mainly to support AWS. I think that view is outdated. The logistics network has become more regionalized and efficient, delivery speeds have improved, and advertising adds high-margin monetization to traffic that retail already generates.

This changes the consolidated economics. Every improvement in fulfillment cost per unit has enormous leverage because the retail base is so large. Advertising can then raise the profit earned from each customer interaction without requiring a separate physical distribution network.

I would not value Amazon retail like a conventional retailer. It is closer to a logistics platform plus marketplace plus advertising network, all sharing customer data and infrastructure.

Advertising is the quiet compounding asset

Amazon advertising does not receive the same attention as AWS because it lacks the spectacle of AI infrastructure. But its economics can be exceptional. A seller already operating on Amazon has a strong incentive to pay for visibility at the moment a customer is ready to purchase.

This creates advertising demand with unusually high commercial intent. Unlike a social platform trying to infer what a user might want, Amazon often knows what the user is actively searching to buy.

The incremental margin can therefore be attractive, and advertising helps subsidize the broader commerce ecosystem. In my valuation, this business deserves a higher multiple than the physical retail operation.

Why the capex can still be rational

Amazon has lived through this pattern before. The company repeatedly spent ahead of visible demand on fulfillment centers, Prime logistics and AWS capacity. Those investments often depressed near-term free cash flow and looked excessive until the infrastructure became difficult for competitors to replicate.

AI infrastructure may be another version of that strategy. Power access, data-center sites, networking design, custom silicon and global cloud regions create barriers that cannot be reproduced quickly.

If AI becomes a foundational layer of enterprise software, the providers that built capacity early may earn returns for a decade. Underinvesting could be more dangerous than overinvesting.

But history is not a guarantee. The scale of current AI spending is much larger, and the technology cycle is faster. Expensive hardware can become economically obsolete before the physical data center has reached the end of its useful life.

Custom chips are Amazon’s defense against Nvidia economics

AWS cannot allow the economics of AI infrastructure to be dictated entirely by one external semiconductor supplier. Trainium and Inferentia are therefore strategically critical even if customers continue demanding Nvidia GPUs.

Custom accelerators can lower Amazon’s cost per unit of compute and give the company more pricing flexibility. They also reduce supply dependence. The better Amazon becomes at designing its own chips, the more bargaining power it has across the supply chain.

This does not mean Amazon needs to replace Nvidia. It needs a credible alternative for enough workloads to prevent its gross margin from being permanently squeezed by external hardware costs.

My valuation framework

I value Amazon using a sum-of-the-parts logic rather than a single P/E. AWS deserves the highest multiple because of its growth, margins and strategic position. Advertising deserves a premium multiple because it is capital light. Retail deserves a lower multiple because it is operationally intensive but has improving economics.

In my bear case, AI capex remains elevated while AWS growth falls back toward the low 20s, depreciation compresses cloud margins and retail efficiency stalls. I would value the equity around $1.6–1.9 trillion.

In my base case, AWS compounds in the high 20s for several years, AI capacity is absorbed, advertising remains strong and retail margins improve gradually. I can justify approximately $2.4–2.8 trillion.

In my bull case, AWS sustains growth above 30% longer than expected, custom silicon protects economics and generative AI creates a new wave of enterprise cloud migration. Then $3.2–3.6 trillion becomes plausible.

At a share price around the mid-$250s in recent trading, Amazon is no longer priced like a neglected mega-cap. The market is paying for a large part of the AI opportunity. I still see upside in the base case, but not the kind of margin of safety that existed when AWS growth was temporarily decelerating and sentiment was weak.

The balance sheet is not the constraint

Unlike smaller AI infrastructure companies, Amazon does not need friendly capital markets to continue investing. Its operating cash flow base is immense, and its access to debt markets is excellent. This removes one of the largest risks in capital-intensive growth.

The constraint is internal return discipline. Management can afford to spend too much. Investors therefore have to judge the productivity of the spending rather than simply celebrate Amazon’s ability to finance it.

What I will watch in the next four quarters

First, AWS growth relative to capex growth. If cloud revenue continues accelerating while capex remains high, the investment case strengthens. If capex rises and AWS decelerates, I become more cautious.

Second, AWS operating margin after depreciation catches up with the new asset base. Third, external demand for Trainium and Inferentia. Fourth, North American retail margin. Fifth, advertising growth.

I will also separate recurring operating profit from mark-to-market gains in AI investments. A rising Anthropic valuation can be valuable, but it should not distract from the economics of Amazon’s own businesses.

What could break the thesis

The biggest risk is overbuilding. Data-center capacity takes time to construct and can arrive after demand has shifted. A second risk is power. The AI cycle is increasingly constrained by electricity, grid connections and cooling rather than chips alone.

A third risk is competition. Microsoft and Google are also investing aggressively, while specialized AI clouds compete for high-end workloads. A fourth is technology obsolescence. If hardware efficiency improves faster than expected, yesterday’s expensive capacity may become less valuable.

Finally, there is the familiar valuation risk. Amazon can execute well and still underperform if investors have already discounted the next five years of success.

My conclusion

Amazon’s second quarter strengthens my conviction in AWS and simultaneously raises my standard for capital allocation. Thirty-seven percent AWS growth at this scale is exceptional. A nearly 40% AWS operating margin is exceptional. More than $160 billion of trailing operating cash flow is exceptional.

Negative trailing free cash flow is the reminder that none of those numbers exist in isolation. Amazon is reinvesting at a rate that would be impossible for almost any other company. The investment can create enormous value, but shareholders need proof that the returns remain high.

I remain constructive on Amazon. I would rather own a company investing aggressively into visible demand than one protecting near-term free cash flow by underbuilding a strategic platform. But I would not ignore the cost of that ambition. The next phase of the stock will be determined by the return on AI capital, not by the size of the AI budget.

For a related analysis of the financing intensity across AI infrastructure, see my CoreWeave deep dive.

Sources and data date

Data date: 20 August 2026. Market values and share prices move continuously. Scenario values are my own framework.

• Amazon Q2 2026 results
• Reuters: AI investment gains and hyperscaler earnings
• Reuters: hyperscaler AI capex and cloud demand

This article reflects my personal analysis and is not investment advice.